Inspiration
Skin cancer is a very serious medical condition which affects the lives of many people even including members of our HackGT team. We personally know people who have been affected by these harmful skin diseases, and some of us are highly susceptible to these diseases ourselves. By creating an application that provides easy access and fast results, we strive to provide the ability for anyone to have reliable information at their disposal regarding skin cancer.
What it does
Skin-tect aims to provide tools and education about how artificial intelligence (AI) can be used to detect skin cancer, while raising awareness about current bias and issues with existing skin lesion detection AI, and working to offer healthcare providers with easy tools to reduce bias in the data and improve accuracy of predictions. We implemented functionality for an everyday user to be able to use our AI to detect potentially dangerous skin lesions, while making sure to provide additional healthcare and technological information. We have also set up a page that will backed by a database, so that healthcare providers can input their medical licensing information to be verified before creating an account to upload images of skin lesions that they diagnose to expand our dataset, reduce bias in AI, and continue to improve the accuracy of the AI’s analysis.
How we built it
Our website is powered by an AI that we created. The model is built on tensorflow and is cross-trained from the Xception model, which has over 22 million parameters. Using the imagenet weights as a base, the model was further trained on the HAM10000 dataset – a dataset of over 10,000 images of different types of skin lesions, including several cancerous and benign classifications. The website was initially modeled entirely on Figma, and from there was converted into html and css files to create the website. Finally we connected the AI model such that users can utilize it through our website.
Challenges we ran into
For the front-end we experienced challenges with bridging the gap between developmental mock-up tools such as Figma and actual code web development. Since none of us had much experience with html or css, learning how to create our website in such a short time period was particularly difficult. For the back-end, we ran into many complications trying to link our trained model with our front-end. Trying to receive the image and generate a prediction was more challenging than we thought it would be. We also had planned to implement a backing database, but we ended up running out of time so we had to scrap the idea.
Accomplishments that we're proud of
We’re proud of creating a respectably accurate model in such a short period of time. We’re also proud of succeeding in connecting front-end and back-end development to create a website that functions to perform the tasks we set out to achieve. Notably, implementing the AI algorithm to be usable from a web interface was something we’re especially proud of.
What we learned
None of us had ever really done web development before. Diving into a full-featured framework like Django was a bit intimidating. With some perseverance, however, we managed to work out many of the kinks in our system to create a fully functioning prototype. We have all now advanced our understanding of web dev, html, and css considerably.
What's next for Skin-tech
The next step for Skin-tech is to attach the database to the web app. This would allow us to keep track of all new data submitted by medical professionals as well as properly address login and account creation. Beyond that, the model we used can definitely be improved. Gathering more data, creating more synthetic data, and better neural network design can help improve the accuracy here.
Log in or sign up for Devpost to join the conversation.